Overview

Natural language interfaces represent a fundamental shift in human-computer interaction, evolving from Dharmesh's 20-year-old concept of "Ingenisoft" to today's AI agents that can translate unstructured human intentions into structured actions, with autonomous multi-agent systems being the next evolutionary step.

Dharmesh defines an agent as "AI-powered software that accomplishes a goal", with various classifications (autonomous vs. non-autonomous, deterministic vs. non-deterministic, synchronous vs. asynchronous), and believes we're entering "the year of agents" where specialized AI workers will form networks similar to professional human networks.

The Agent.ai platform (with 1.3 million users and 3,000 agent builders) aims to create a marketplace of specialized agents that can interact like "Lego building blocks," each with REST APIs, enabling developers to compose functionality across a network while providing flexible model selection to optimize for cost and performance.

AI is transforming software development by dramatically reducing the cost of code refactoring and generation, though this creates risks of unnecessary product complexity; the future value of engineers will come from communication skills, customer understanding, and the ability to effectively collaborate with AI tools.

The concept of "work as a service" (where software performs actual work rather than being a tool) is emerging, with potential for outcomes-based pricing models in industries with objectively measurable results, while cross-agent memory sharing represents the "next frontier" for creating personalized user experiences with appropriate privacy controls.

Content

Dharmesh Shah's AI Journey and Vision

* Dharmesh Shah is the founder of Agent.ai and co-founder of HubSpot, with a long-standing interest in natural language interfaces for software. * 20 years ago, he conceived "Ingenisoft" - a concept for a natural language interface to business software, initially envisioning email as an offline interface for business systems. * He believed traditional software interfaces are unintuitive, requiring users to translate mental models into clicks and swipes. * Post-ChatGPT, he developed ChatSpot as an early exploration of natural language interactions. * He sees natural language interfaces as a breakthrough, allowing users to directly express intentions and appreciates AI's ability to convert unstructured input into structured actions. * Dharmesh views chat interfaces as just a first step, with the next significant advancement being software that can perform multi-step tasks autonomously. * He sees agents as the next evolutionary stage of software interaction.

Defining and Understanding Agents

* Dharmesh broadly defines an agent as "AI-powered software that accomplishes a goal." * Agents can be classified along multiple dimensions: - Autonomous vs. non-autonomous - Deterministic vs. non-deterministic workflow - Synchronous vs. asynchronous operation - Different interaction modes (e.g., chat agent, workflow agent)

* Early implementations like Baby AGI and AutoGBT were conceptual thought experiments that assumed reasoning and execution capabilities that didn't yet exist. * Current focus is on practical, commercially viable agent applications.

* Factors enabling current agent development include: - Better AI models - More reliable tool use - Dropping costs - Faster inference - Increased model diversity - Multi-agent research - Reinforcement learning fine-tuning

* The term "agent" etymologically means something that acts on behalf of someone. * Dharmesh explores the fundamental definition of an "agent," drawing parallels to biological concepts like single-celled organisms. * He proposes viewing everything as potentially an "agent," with tools potentially being considered atomic agents. * He believes an agent must, at some level, use AI to be considered an agent. * Dharmesh suggests this is "the year of agents," with multi-agent systems likely emerging next. * He draws analogies between biological complexity (single cell → multi-cell → organs → complete organism) and potential technological agent development.

Knowledge Representation and Graph Databases

* The discussion shifts to graph data structures and their potential in AI and knowledge management. * Graphs represent an alternative to traditional database structures like relational databases and are potentially better for representing knowledge for LLMs and AI. * Graphs are more discoverable and observable compared to embeddings and can provide more contextual understanding than simple vector embeddings.

* Current RAG (Retrieval Augmented Generation) methods lose context through chunking. * Graph-based knowledge stores might yield better results by preserving broader context. * There's potential to rank nodes/contributors based on authority or relevance.

* Graphs offer a different structural approach compared to tables and columns. * PageRank inspired potential for "node rank" algorithms in knowledge graphs. * Graph databases are gaining traction in AI, especially for conversation memory and document retrieval.

* However, graph databases can become overly complex with a risk of "over-engineering" solutions. * Most ML practitioners remain skeptical of knowledge graphs.

* There's a pragmatic view that a data store can be useful even if its underlying mechanism is not fully understood. * Graph databases offer unique traversal capabilities that relational databases cannot efficiently replicate. * Representing complex structures (like social graphs) in relational databases often doesn't scale or work effectively.

Engineering Philosophy and Decision-Making

* Dharmesh discusses key engineering considerations including: - Assessing the cost of doing something "the right way" - Evaluating potential future scenarios and their probabilities - Considering the return on engineering effort

* He expresses a preference for under-engineering over over-engineering because: - Technical debt is more predictable and manageable - If anticipated use cases never materialize, time/resources are saved - Good engineers can identify when doing something "right" has minimal additional cost

* His decision-making approach involves: - Thinking like a business person: calculating expected value - Assigning probabilities to potential future scenarios - Balancing between current needs and potential future requirements - Aiming to create solutions with compounding value rather than accumulating technical debt

AI's Impact on Software Development

* The cost of refactoring and fixing code is trending towards zero due to emerging AI technologies. * AI code generation models are making it easier to quickly add features and create code. * There's concern that low-effort feature addition might lead to unnecessary product complexity. * There's a risk of being "promiscuous" with product extensions without careful consideration. * AI models may not inherently understand the long-term consequences of adding complexity.

* The future progression of AI in software development may go from code autocomplete to full function generation to entire app/platform generation. * A potential future scenario includes AI generating entire software companies, including platforms, monetization, and go-to-market strategies.

* Junior engineers are not necessarily becoming obsolete. Their value will come from: - Communication skills - Customer interaction - Understanding fundamental software principles - Ability to work with AI tools effectively

* Learning programming is less about syntax and more about understanding how systems work. * Coding education provides fundamental problem-solving and logical thinking skills.

Multimodal Context Protocols and Standards

* Dharmesh highlights Multimodal Context Protocols (MCP) as a significant development: - He views MCP as an important standard in the AI ecosystem - MCPs are seen as accessible and understandable for engineers - They enable easier integration and communication between different tools and systems

* For future AI and agent systems, he: - Anticipates the need for agent collaboration and discovery mechanisms - Believes MCPs (or similar standards) will be crucial for enabling inter-system communication - Sees potential for diverse MCP clients beyond current chat-based applications

* Comparing with OpenAPI: - OpenAPI provides structured, machine-readable API definitions - MCPs are more use-case specific and potentially more flexible

* Regarding creating standards, Dharmesh: - Doesn't feel confident creating standards himself - Prefers consuming and combining existing technologies creatively - Lacks credibility to drive widespread standard adoption

* He proposes an "Open Graph" standard concept: - Aims to give individuals control over their social/professional data - Would allow personal data portability across platforms - Envisioned as a public benefit, non-profit initiative

* On data ownership: - Current social platforms (Meta, LinkedIn) have closed data ecosystems - There's a desire for individuals to control and easily export their own data - The concept is inspired by protocols like AT Proto (behind Blue Sky) - The challenge with data portability standards is getting average users to care about the issue

Privacy, Data Sharing, and AI Agents

* Dharmesh discusses his personal willingness to trade privacy for productivity or benefits. * He highlights LinkedIn's strict stance on data usage and protection. * He mentions a Supreme Court case where LinkedIn successfully defended against data scraping.

* For AI agents and future teams, he: - Predicts the emergence of "hybrid teams" combining human workers and AI agents - Suggests agents will eventually have professional networks similar to LinkedIn - Proposes agent.ai as a potential "LinkedIn for agents" with profiles, connections, and posting capabilities

* Key conceptual insights include: - Agents should be treated as tools/teammates, not anthropomorphized - Agents should self-disclose as software - Future professional networking might include digital/AI entities - There's a need for structured ways to evaluate and understand AI agents

Agent.ai Platform Development

* Dharmesh discusses the development of agent.ai, a platform for creating digital agents/workers: - 1.3 million users - 3,000 people have built agents - About 1,000 agents have been published

* The key concept is creating a network of specialized agents that can interact and compose functionality: - Example: A domain valuation agent that can be used by other agents - Example workflow: A brand-building agent could use a domain valuation agent to find potential startup names and domain prices

* Technical challenges being explored include: - How to manage and discover agents across a network (using MCP - Master Control Program) - Limitations of how many tools/agents an LLM can effectively use - Developing an intelligent routing/selection mechanism to choose relevant agents for a given task (described as "RAG for tools")

* Platform goals include: - Making agent creation easier through a low-code platform - Enabling developers to access and combine agents like "Lego building blocks" - Creating a community-driven ecosystem of specialized digital workers

* Current offering includes an MCP server that provides: - Access to multiple models via a single API key - Internally built agents with specific functionalities - Free access (currently funded by Dharmesh)

Design Philosophy and UI Evolution

* Dharmesh initially chose a low-code, deterministic approach due to: - Lack of advanced reasoning models at the time - Belief that known steps should be explicitly defined rather than left to chance - No perceived upside in randomizing known processes

* For UI and agent interaction models, he: - Predicts a shift from current chatbot models to hybrid interaction approaches - Envisions asynchronous interactions more similar to workplace collaboration - Notes that current agent UIs are extremely basic, resembling early HTML form interfaces

* He's exploring code generation for UI within agent workflows: - Describing desired UI - Generating code iteratively - Refining through prompt-based interactions - Caching final generated code to reduce inference costs

* Dharmesh appreciates Python's role as an AI programming language, viewing it as a useful "common denominator" for both humans and AI. * Python's ability to provide runtime interpreter access enables solving problems not originally trained for. * He sees potential for AI to generate and innovate UI beyond current interaction primitives (checkboxes, radio buttons).

* Emerging concept of "generative UI" where: - UIs can be generated on-the-fly - Generated interfaces can be "pinned" and reused - Future AI might create novel interaction primitives * He proposes a model of AI agents filling "Mad Libs" style generated forms.

* He introduced the term "AI shoring" - moving roles/work to AI. * He suggested agents could be evaluated through "proof of work" and running simulations, unlike human professional network evaluations.

Agent.ai Platform Features and Business Model

* Every agent on the platform automatically has a REST API that can be called. * The goal is to enable easy evaluation and selection of agents for specific tasks. * The proposed model includes: - Ability to test agents multiple times before paying - Audit trails of agent usage - Human ratings and reviews (currently averaging 4.1/5 stars)

* Current challenge: Organizations talk about evaluations but rarely implement robust testing. * Suggested approach: Generate real-world test scenarios for AI agents (e.g., software engineering tasks). * Comparison to traditional hiring practices like take-home projects.

* Regarding OpenAI and GPT Store: - Sees OpenAI's development of agent/custom GPT platforms as "inevitable" - Not directly competing with OpenAI, but filling a specific need - Agent.ai originated from personal productivity projects

* Technical design philosophy: - Desire to "straddle models" - use different AI models for different task steps - Flexibility in model selection based on specific requirements (e.g., writing, image generation)

* Key motivation: - Create a framework for building AI-driven personal productivity tools - Provide infrastructure for more flexible, modular AI agent development

Model Selection and Pricing Models

* Dharmesh advocates for a flexible approach to AI model selection, allowing users to mix and match models. * Users tend to choose the largest/most powerful model by default, which can be unnecessarily expensive. * There's potential for significant cost reduction by routing to lower-cost models without sacrificing output quality. * The goal is to develop "model routing" - helping users pick the best model at runtime.

* On marketplaces and market efficiency: - He is passionate about creating efficient markets in AI - Inspired by Chai AI's marketplace model that allows model providers to compete - Believes inefficient markets exist due to: - Lack of knowledge between buyers and sellers - Absence of trust mechanisms - Difficulty in pricing/determining fair market value

* He introduces the concept of "work as a service" - where software performs the actual work, unlike traditional software as a service (SaaS). * This contrasts with traditional models where humans consume the software.

* The discussion shifts to "results as a service" or outcomes-based pricing models in various industries: - Currently most popular in customer support, where pricing is based on resolved tickets - Customer support has clear advantages for this model: - Objective measurement of resolution - Consistent economic value per ticket - Semi-objective metrics like Net Promoter Score or CSAT

* Challenges with outcomes-based pricing: - Many industries lack clear, objective outcome measurements - Value per outcome can vary dramatically across different sectors - Subjective fields like design pose significant measurement challenges: - Quality of output isn't fully controllable - Requires multiple iterations - Customer satisfaction is hard to standardize

* Using logo design as a specific example: - 99designs as a case study - More iterations (200 vs 20) increases likelihood of finding a satisfactory design - Designers are incentivized by commitment to selecting a winner

* Future outlook: - Some markets are likely to move toward results-driven business models - Key considerations include: - Ability to objectively measure outcomes - Pricing consistency - Low variability in results

Attribution, Blockchain, and Software Engineering

* The discussion covers attribution challenges across industries: - Current systems lack meaningful ways to track and attribute value - Companies often lack incentives to share data that could enable better attribution - Google has progressively reduced data transparency (e.g., keyword information)

* On Web3 and blockchain potential: - The speakers see potential in blockchain's fundamental principles - Blockchain could enable better tracking, fractionalization of digital assets, and verifiable audit logs - Previous Web3 efforts were undermined by "crypto bros" and speculative NFT markets - Blockchain has limited but meaningful use cases

* Regarding software engineering and AI: - Discussion on whether software engineering will remain a "work as service" model - Dharmesh is bullish on engineers' long-term economic value - AI and co-generation tools are expected to: - Increase total economic value possible through software - Provide engineers with more powerful problem-solving tools - Contrary to fears about AI replacing engineers, he believes engineer value will actually increase - Engineers will have enhanced capabilities to solve a broader range of problems

Memory and Agent Development

* Dharmesh encourages people, regardless of technical background, to learn and use AI agents. * There's an emphasis on being a "techno optimist" and not being scared of AI technology. * Engineers are still in early stages of developing the "agent's stack".

* Long-term memory is considered the "next frontier" for agents and agent networks. * He is particularly interested in "cross-agent memory" - the ability for different agents to share learned context about a user. * A key goal is creating user utility through shared, contextual memory across agents.

* On authentication and privacy considerations: - Current authentication models (like OAuth 2.0) are seen as too coarse and limited - He rejects the term "privacy first" as overly absolute - Selective memory sharing is crucial - agents should only access relevant, authorized information

* His future vision includes: - Developing networks of agents with shared, trustable memory - Creating platforms where agent builders can leverage collective user knowledge - Solving fundamental challenges around memory, authentication, and context sharing

* Regarding data privacy and sharing: - Discussion about carefully sharing data subsets with third-party services - Proposal for a "trusted intermediary" that can control data sharing with fine-grained permissions - Example: Sharing email data with strict controls like limiting number of emails or specific labels

* Technology and memory observations: - Critique of current OAuth and data sharing limitations - Emerging interest in AI memory hierarchies (semantic, episodic, background processing) - Emerging concept that AI systems should "sleep" to process memories, similar to human cognition

* Potential AI product opportunities: - Need for better email AI integration and semantic search - Potential for shared team/group agent memory spaces - Platforms should provide selective conversation/memory sharing capabilities

* Memory technologies mentioned: - Mem Zero - MemGPT - ZEP (uses graph database) - LangMem from LangGraph

The Agent Network — Dharmesh Shah